Re: [PATCH 1/4] ARC: add barriers to futex code

[Will Deacon <will.deacon@arm.com>]

On Thu, Aug 06, 2015 at 02:48:26PM +0100, Peter Zijlstra wrote:
> On Thu, Aug 06, 2015 at 06:05:20PM +0530, Vineet Gupta wrote:
> > The atomic ops on futex need to provide the full barrier just like
> > regular atomics in kernel.
> > 
> > Also remove pagefault_enable/disable in futex_atomic_cmpxchg_inatomic()
> > as core code already does that
> 
> Urgh, and of course tglx just left for holidays :-)

Damn, he's really missing out on this!

> > +++ b/arch/arc/include/asm/futex.h
> > @@ -20,6 +20,7 @@
> >  
> >  #define __futex_atomic_op(insn, ret, oldval, uaddr, oparg)\
> >  							\
> > +	smp_mb();					\
> >  	__asm__ __volatile__(				\
> >  	"1:	llock	%1, [%2]		\n"	\
> >  		insn				"\n"	\
> > @@ -40,12 +41,14 @@
> >  							\
> >  	: "=&r" (ret), "=&r" (oldval)			\
> >  	: "r" (uaddr), "r" (oparg), "ir" (-EFAULT)	\
> > -	: "cc", "memory")
> > +	: "cc", "memory");				\
> > +	smp_mb();					\
> >  
> 
> 
> So:
> 
>  - alhpa: only has the first smp_mb(), suggesting RELEASE
>  - arm: only has the first smp_mb(), suggesting RELEASE
>  - arm64: has store-release + smp_mb(), suggesting full barriers

I'd be ok relaxing that to smp_mb() but I don't think I'm brave enough
to go all the way to an STLXR. You can lose SC if you combine explicit
barrier instructions with the acquire/release instructions and I dread
to think what userspace is doing...

>  - MIPS: has LLSC_MB after, suggesting ACQUIRE

Yikes, so there's a fun semantic difference there. Maybe we should go
look at glibc (which only uses one of the futex ops in pthread_cond_wait
iirc).

>  - powerpc: lwsync before, sync after, full barrier
> 
> x86 is of course boring and fully ordered
> 
> Looking at the usage site of futex_atomic_op_inuser(), that's in
> futex_wake_op() which might suggest RELEASE is indeed sufficient.
> 
> Which leaves me puzzled on MIPS, but what do I know.
> 
> At the very least this patch isn't wrong, fully ordered is sufficient.

Agreed.

Will